M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images

Hongyi Wang, Xiuju Du, Jing Liu, Shuyi Ouyang, Yen-Wei Chen, Lanfen LinAAAI 2025 · DOI

Abstract

Spatial Transcriptomics (ST) provides critical insights into the tumor microenvironment by mapping gene expression to specific tissue locations. However, the high cost and technical complexity of ST limit its clinical accessibility. Digital pathology images, or Whole Slide Images (WSIs), are ubiquitous and cost-effective, but predicting gene expression from morphology remains challenging due to the multi-scale nature of tissue architecture. We introduce M2OST, a many-to-one regression Transformer that predicts spatial gene expression by simultaneously analyzing multiple magnification levels. M2OST achieves state-of-the-art accuracy across multiple datasets while maintaining a lightweight profile (6.81M parameters) and delivering 100x faster inference than competing methods.

Introduction

Understanding the spatial distribution of gene expression is vital for characterizing tumor microenvironments and identifying biomarkers. While Spatial Transcriptomics (ST) enables this, its prohibitive cost has led researchers to explore computational methods for predicting transcriptomic profiles directly from hematoxylin and eosin (H&E) stained digital pathology slides.

Overview of Whole Slide Imaging (WSI) and Spatial Transcriptomics (ST) mapping.

Existing models often suffer from "tunnel vision," focusing on single-scale image patches and ignoring the hierarchical context inherent in pathology "pyramids." M2OST addresses this by mimicking the diagnostic workflow of pathologists, who integrate information across various magnifications to understand tissue morphology.

Methods

M2OST utilizes a "many-to-one" scheme where three patches from different magnification levels (e.g., 5x, 10x, and 40x) centered on the same spatial coordinate are processed together to predict a single gene expression vector.

Architectural Components

The model is built on a Transformer backbone with several specialized modules designed for efficiency and multi-scale integration:

  • Deformable Patch Embedding (DPE): Generates tokens that differentiate between the high-resolution central target spot and the surrounding coarse-grained context.
  • Decoupled Feature Extraction: To optimize efficiency and avoid the quadratic complexity of standard Transformers, feature learning is split into three stages:
  • Intra-Level Transformer (ITMM): Captures spatial dependencies within each magnification level.
  • Cross-Level Transformer (CTMM): Uses cross-attention to fuse information across different scales, allowing lower-magnification context to inform high-resolution details.
  • Cross-Channel Mixer (CCMM): Refines features across channels using a squeeze-and-excitation mechanism.
The M2OST architecture featuring multi-scale input and decoupled feature extraction modules.

Experimental Results

M2OST was evaluated on three public datasets: Human Breast Cancer (HBC), HER2-positive Breast Tumor (HER2+), and Human Cutaneous Squamous Cell Carcinoma (cSCC).

Performance and Efficiency

The model consistently outperformed both CNN-based and Transformer-based baselines in Pearson Correlation Coefficient (PCC), reaching 48.07% on HBC and 50.50% on cSCC.

Beyond accuracy, M2OST is exceptionally efficient:

  • Parameter Count: 6.81 million, compared to 675 million for the competing Hist2ST model.
  • Computational Load: 2.24 GFLOPs.
  • Inference Speed: 100x faster than the state-of-the-art iStar model.

Discussion and Conclusion

The success of M2OST underscores the importance of multi-scale context in computational pathology. By decoupling intra-scale and cross-scale learning, the model achieves high performance without the computational overhead typical of large-scale Transformers.

While M2OST currently focuses on spot-level predictions, future work could integrate global slide-level dependencies to further refine spatial maps. The architecture's flexibility also suggests potential for multi-modal integration, extending beyond transcriptomics to other biological data types.

Summary updated

BibTeX
@inproceedings{wang2025m,
  author = {Hongyi Wang and Xiuju Du and Jing Liu and Shuyi Ouyang and Yen-Wei Chen and Lanfen Lin},
  title = {{M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i7.32830},
}
Figure